Datasets:
Spatial Soft Pastry005 HDF5 Data Format
Dataset:
/vepfs-C区/visuotactile/Tabero/formal_collections/spatial_soft_pastry005_10tasks_50each_gripperjitter_replay_format_20260624
Overall Layout
There are 10 task HDF5 files:
libero_spatial/libero_spatial_task*/replayed_demos/*_replayed_demo.hdf5
Each HDF5 has:
/data/demo_0
/data/demo_1
...
/data/demo_49
So each task has 50 demos. The first dimension T below is the episode frame count and can vary across demos.
Demo Attributes
Stored on each /data/demo_* group:
| Attribute | Meaning |
|---|---|
asset_name |
Soft object name, here soft_pastry005. |
demo_id |
New or collection-time demo id. |
original_demo_id |
Original script expert demo id. |
task_suite |
LIBERO suite, here libero_spatial. |
task_id |
Task id, 0..9. |
soft_task_kind |
Soft collection task name, e.g. libero_spatial_soft_task0. |
task |
Isaac/Tabero environment id used during collection. |
language |
Natural-language instruction used for the demo. |
success |
Whether the script expert collection succeeded. |
failure_reason |
Failure reason if success=False; empty for success. |
num_frames |
Number of frames in this demo. |
num_samples |
Same frame/sample count used by replay-format tools. |
metadata_json |
Extra metadata, including frames, gripper_close_norm, mode, place_in_target. |
source_hdf5 |
Raw source HDF5 path before replay-format assembly. |
Action And Observation Fields
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
actions |
(T, 13) |
float32 |
Training action aligned with replay format: [x, y, z, axis_angle_x, axis_angle_y, axis_angle_z, gripper, force_left_x, force_left_y, force_left_z, force_right_x, force_right_y, force_right_z]. This is the action OpenPI uses with state = obs/eef_pose and delta = action - state style preprocessing. Axis-angle branch was aligned to avoid ±pi discontinuity. |
obs/actions |
(T, 8) |
float32 |
Raw robot command before 13D conversion: [x, y, z, quaternion(4), gripper]. Mostly equivalent to soft_extras/actions_raw. |
obs/eef_pose |
(T, 7) |
float32 |
End-effector state used as policy state: [x, y, z, axis_angle_x, axis_angle_y, axis_angle_z, gripper]. This is numerically aligned with /actions rotation convention. |
obs/arm_joint_pos |
(T, 7) |
float32 |
Franka arm joint positions, arm joints only. |
obs/gripper_pos |
(T, 2) |
float32 |
Two gripper finger joint positions. |
obs/gripper_width |
(T,) |
float32 |
Scalar gripper opening width. |
obs/gripper_close_norm |
(T,) |
float32 |
Normalized close value used by the script expert for pastry gripping. This records the gripper jitter value/process in normalized form. |
obs/gripper_net_force |
(T, 1, 2, 3) |
float32 |
Net force observation for tactile/force policy input. The last two dimensions are left/right gripper side and 3D force vector. |
obs/gripper_marker_motion |
(T, 2, 2, 99, 2) |
float32 |
Marker-motion tactile observation. It stores left/right tactile marker displacement fields; each marker has 2D image-plane motion. The 99 dimension is marker count. |
obs/fem_bbox_dims |
(T, 3) |
float32 |
Soft pastry FEM bounding-box size over time, [dx, dy, dz]. Used for deformation/safety diagnostics. |
obs/fem_deformation_max |
(T,) |
float32 |
Per-frame maximum FEM deformation metric for the soft pastry. Used by the safety check. |
obs/fem_deformation_rms |
(T,) |
float32 |
Per-frame RMS FEM deformation metric for the soft pastry. Used by the safety check. |
Soft Extras
These fields duplicate or preserve soft-specific collection signals in a separate namespace so replay-compatible common fields remain stable.
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
soft_extras/actions_13d |
(T, 13) |
float32 |
Soft collection copy of the final 13D action. Should match /actions. |
soft_extras/actions_raw |
(T, 8) |
float32 |
Raw 8D action before conversion to replay-style 13D. |
soft_extras/arm_joint_pos |
(T, 7) |
float32 |
Soft collection copy of arm joint positions. |
soft_extras/gripper_width |
(T,) |
float32 |
Soft collection copy of scalar gripper width. |
soft_extras/gripper_close_norm |
(T,) |
float32 |
Soft collection copy of normalized close value. |
soft_extras/gripper_net_force |
(T, 1, 2, 3) |
float32 |
Processed net force used for observation/training. |
soft_extras/gripper_net_force_raw |
(T, 1, 2, 3) |
float32 |
Raw net force before any fallback/processing. |
soft_extras/gripper_net_force_source |
(T,) |
int8 |
Source flag for net force values. Used to distinguish measured/raw/fallback force source. |
soft_extras/gripper_marker_motion |
(T, 2, 2, 99, 2) |
float32 |
Soft collection copy of marker-motion tactile field. |
soft_extras/fem_bbox_dims |
(T, 3) |
float32 |
Soft collection copy of FEM bbox dimensions. |
soft_extras/fem_deformation_max |
(T,) |
float32 |
Soft collection copy of max deformation metric. |
soft_extras/fem_deformation_rms |
(T,) |
float32 |
Soft collection copy of RMS deformation metric. |
Dynamic State Fields
These fields store simulator state at every frame. Pose convention follows Isaac/Tabero root state convention: root pose is position plus quaternion, and root velocity is linear plus angular velocity.
Robot And Articulations
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
states/articulation/robot/joint_position |
(T, 9) |
float32 |
Robot articulation joint positions: 7 arm joints plus 2 gripper joints. |
states/articulation/robot/joint_velocity |
(T, 9) |
float32 |
Robot articulation joint velocities. |
states/articulation/robot/root_pose |
(T, 7) |
float32 |
Robot root pose. |
states/articulation/robot/root_velocity |
(T, 6) |
float32 |
Robot root linear/angular velocity. |
states/articulation/flat_stove_1/joint_position |
(T, 1) |
float32 |
Flat stove articulation joint position in the spatial scene. |
states/articulation/flat_stove_1/joint_velocity |
(T, 1) |
float32 |
Flat stove articulation joint velocity. |
states/articulation/flat_stove_1/root_pose |
(T, 7) |
float32 |
Flat stove root pose. |
states/articulation/flat_stove_1/root_velocity |
(T, 6) |
float32 |
Flat stove root velocity. |
states/articulation/wooden_cabinet_1/joint_position |
(T, 3) |
float32 |
Wooden cabinet articulation joint positions. |
states/articulation/wooden_cabinet_1/joint_velocity |
(T, 3) |
float32 |
Wooden cabinet articulation joint velocities. |
states/articulation/wooden_cabinet_1/root_pose |
(T, 7) |
float32 |
Wooden cabinet root pose. |
states/articulation/wooden_cabinet_1/root_velocity |
(T, 6) |
float32 |
Wooden cabinet root velocity. |
Deformable Object
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
states/deformable_object/soft_pastry005/root_pos_w |
(T, 3) |
float32 |
Soft pastry root/world position. |
states/deformable_object/soft_pastry005/nodal_pos_w |
(T, 115, 3) |
float32 |
FEM node world positions for the soft pastry. 115 is the number of FEM nodes in this asset. |
states/deformable_object/soft_pastry005/bbox_min_w |
(T, 3) |
float32 |
World-space minimum corner of the pastry FEM bounding box. |
states/deformable_object/soft_pastry005/bbox_max_w |
(T, 3) |
float32 |
World-space maximum corner of the pastry FEM bounding box. |
Rigid Scene Objects
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
states/rigid_object/akita_black_bowl_1/root_pose |
(T, 7) |
float32 |
Akita black bowl root pose in the spatial scene. |
states/rigid_object/akita_black_bowl_1/root_velocity |
(T, 6) |
float32 |
Akita black bowl root velocity. |
states/rigid_object/cookies_1/root_pose |
(T, 7) |
float32 |
Cookies object root pose. |
states/rigid_object/cookies_1/root_velocity |
(T, 6) |
float32 |
Cookies object root velocity. |
states/rigid_object/glazed_rim_porcelain_ramekin_1/root_pose |
(T, 7) |
float32 |
Ramekin root pose. |
states/rigid_object/glazed_rim_porcelain_ramekin_1/root_velocity |
(T, 6) |
float32 |
Ramekin root velocity. |
states/rigid_object/plate_1/root_pose |
(T, 7) |
float32 |
Plate root pose. |
states/rigid_object/plate_1/root_velocity |
(T, 6) |
float32 |
Plate root velocity. |
states/rigid_object/soft_pastry005_tactile_proxy/root_pose |
(T, 7) |
float32 |
Tiny rigid tactile proxy pose for the soft pastry. |
states/rigid_object/soft_pastry005_tactile_proxy/root_velocity |
(T, 6) |
float32 |
Tiny rigid tactile proxy velocity. |
Initial State Fields
initial_state/* mirrors states/*, but only stores the initial frame with shape first dimension 1. These fields are used for reset/replay and for reproducing the starting scene.
Robot And Articulations
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
initial_state/articulation/robot/joint_position |
(1, 9) |
float32 |
Initial robot joint positions. |
initial_state/articulation/robot/joint_velocity |
(1, 9) |
float32 |
Initial robot joint velocities. |
initial_state/articulation/robot/root_pose |
(1, 7) |
float32 |
Initial robot root pose. |
initial_state/articulation/robot/root_velocity |
(1, 6) |
float32 |
Initial robot root velocity. |
initial_state/articulation/flat_stove_1/joint_position |
(1, 1) |
float32 |
Initial flat stove joint position. |
initial_state/articulation/flat_stove_1/joint_velocity |
(1, 1) |
float32 |
Initial flat stove joint velocity. |
initial_state/articulation/flat_stove_1/root_pose |
(1, 7) |
float32 |
Initial flat stove root pose. |
initial_state/articulation/flat_stove_1/root_velocity |
(1, 6) |
float32 |
Initial flat stove root velocity. |
initial_state/articulation/wooden_cabinet_1/joint_position |
(1, 3) |
float32 |
Initial wooden cabinet joint positions. |
initial_state/articulation/wooden_cabinet_1/joint_velocity |
(1, 3) |
float32 |
Initial wooden cabinet joint velocities. |
initial_state/articulation/wooden_cabinet_1/root_pose |
(1, 7) |
float32 |
Initial wooden cabinet root pose. |
initial_state/articulation/wooden_cabinet_1/root_velocity |
(1, 6) |
float32 |
Initial wooden cabinet root velocity. |
Deformable Object
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
initial_state/deformable_object/soft_pastry005/root_pos_w |
(1, 3) |
float32 |
Initial soft pastry root/world position. |
initial_state/deformable_object/soft_pastry005/nodal_pos_w |
(1, 115, 3) |
float32 |
Initial FEM node world positions. |
initial_state/deformable_object/soft_pastry005/bbox_min_w |
(1, 3) |
float32 |
Initial pastry FEM bounding-box minimum corner. |
initial_state/deformable_object/soft_pastry005/bbox_max_w |
(1, 3) |
float32 |
Initial pastry FEM bounding-box maximum corner. |
Rigid Scene Objects
| Field | Shape | Dtype | Meaning |
|---|---|---|---|
initial_state/rigid_object/akita_black_bowl_1/root_pose |
(1, 7) |
float32 |
Initial Akita black bowl root pose. |
initial_state/rigid_object/akita_black_bowl_1/root_velocity |
(1, 6) |
float32 |
Initial Akita black bowl root velocity. |
initial_state/rigid_object/cookies_1/root_pose |
(1, 7) |
float32 |
Initial cookies object root pose. |
initial_state/rigid_object/cookies_1/root_velocity |
(1, 6) |
float32 |
Initial cookies object root velocity. |
initial_state/rigid_object/glazed_rim_porcelain_ramekin_1/root_pose |
(1, 7) |
float32 |
Initial ramekin root pose. |
initial_state/rigid_object/glazed_rim_porcelain_ramekin_1/root_velocity |
(1, 6) |
float32 |
Initial ramekin root velocity. |
initial_state/rigid_object/plate_1/root_pose |
(1, 7) |
float32 |
Initial plate root pose. |
initial_state/rigid_object/plate_1/root_velocity |
(1, 6) |
float32 |
Initial plate root velocity. |
initial_state/rigid_object/soft_pastry005_tactile_proxy/root_pose |
(1, 7) |
float32 |
Initial rigid tactile proxy root pose. |
initial_state/rigid_object/soft_pastry005_tactile_proxy/root_velocity |
(1, 6) |
float32 |
Initial rigid tactile proxy root velocity. |
Replay Alignment Notes
- Common replay-compatible training fields are preserved:
actionsobs/actionsobs/eef_poseobs/arm_joint_posobs/gripper_posstates/*initial_state/*
- Soft-specific information is added without replacing common replay fields:
obs/fem_*soft_extras/*states/deformable_object/soft_pastry005/*
/actionsis the 13D training action. The rotation part is axis-angle and is branch-aligned withobs/eef_poseto avoidaction - stateproducing artificial2*pijumps.obs/actionsandsoft_extras/actions_rawpreserve the original 8D absolute command form.
Safety Fields
For safety evaluation, use:
obs/fem_deformation_max
obs/fem_deformation_rms
obs/fem_bbox_dims
The current safety metric checks:
safe = max(obs/fem_deformation_max) <= threshold_max
AND max(obs/fem_deformation_rms) <= threshold_rms
The combined evaluation metric is:
safe_success = success AND safe